Spent a couple hours diving down this rabbit hole, mostly thoughts, ideas around this New Abnormal we live in.

I keep coming back to the idea that we are probably still thinking about AI agents too much like software and not enough like participants inside a system. Reason being is decades were spent building systems around the assumption that a human was somewhere in the loop, and now we are slowly removing that assumption while still relying on a lot of the same infrastructure underneath it.

My pleb spidey sense can't help but feel an agent identity ends up becoming one of those boring sounding problems that turns out to be foundational once this stuff starts operating at real scale.

WWKIRL (what we know is real life): A model may propose or initiate an action, but in a properly designed system, a separately enforced identity and policy layer should determine whether that action is actually authorized to happen.

Then there is a whole other issue with TTFT, that Cisco is discussing in their Day in the Life of a Prompt research, and the interesting part to me is what happens as inference gets faster. Today, network latency is mostly swallowed by the time it takes the model to do the work, but agents are much chatty kathy's than the software we are used to. So one request might turn into tens or even hundreds of calls between an agent, models, tools, and other services. Cisco states as the models get faster, the network starts becoming a larger part of the actual wait. It reminds me a little of infinite loopback scenarios, except now being applied locally across a growing number of agent interactions, could lead to another type of internal attack.

Which makes me wonder, if we play the tape all the way out, whether we eventually end up with protocols designed specifically around agents rather than continuing to bolt agent behavior onto infrastructure designed primarily for humans and traditional applications. Sure, MCP and identity feel like parts of that new stack, but discovery, authorization, routing, payments, state, and low latency communication between agents probably all start converging here too.